Epipolar transfer enables accurate depth from drone video without training
EpiTransfer: Sparse, Training-Free Long-Range Depth Estimation from Temporal Monocular Aerial Frames
RoboticsComputer Vision and Pattern Recognition
Summary
Getting a 3D understanding of a scene is important for things like drone navigation and avoiding obstacles. The authors developed a method that estimates depth from just two camera images taken at different times, without needing any prior training or data. They use the known movement of the camera to create a virtual stereo pair, allowing them to calculate depth with solid accuracy. Their approach works well both indoors and outdoors and often performs better than some popular AI-based depth estimation methods that were not specially trained for drone footage.
What this means in practice
- •For drone developers: Implement long-range depth sensing using only two drone images and known camera movements without collecting training data.
- •For indoor robotics teams: Provide reliable 3D environment mapping from monocular video streams without retraining deep learning models.
Authors
Diksha Aggarwal, Rutvik Dagadkhair, Sanjana Srivastava, Bradley Denby, Kevin Kochersberger
Abstract
Reliable 3D spatial understanding is essential for autonomous navigation, obstacle avoidance, and scene reconstruction. While state-of-the-art learned depth estimation techniques achieve high accuracy in-distribution, they often generalize poorly to novel viewpoints and altitudes. This paper presents a geometrically derived, training-free depth estimation method using epipolar transfer with only two monocular images and camera pose estimates. By leveraging camera motion to synthesize a virtual stereo pair with a freely chosen baseline, our approach transforms temporal correspondence into a stereo triangulation task while mitigating geometric degeneracies inherent to direct two-view triangulation. Validated across outdoor drone flights (to a maximum range of approximately 90\,m) and indoor OptiTrack environments against LiDAR ground truth, the method achieves an indoor AbsRel of 0.092 and $δ< 1.25$ of 0.940, comparable to direct triangulation (AbsRel 0.073) while retaining valid depth over a larger fraction of challenging scenes, and substantially outperforms off-the-shelf learning-based baselines such as ZoeDepth (AbsRel 0.225) and Depth Anything V2 (AbsRel 0.570), which are not trained or fine-tuned for this domain, with no training data required.